Drought propagation time prediction method based on machine learning
A drought propagation time prediction model was constructed through machine learning methods, and the key driving factors were identified using the extreme gradient boosting algorithm. This solved the problem of prediction uncertainty using traditional methods, achieved accurate prediction of the propagation time from meteorological drought to agricultural drought, and supported early warning of agricultural drought and water resource optimization.
Patent Information
- Application Number
- CN202510816627.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-18
- Publication Date
- 2025-10-14
AI Technical Summary
Existing drought propagation time prediction methods lack a practical prediction model, and the prediction results of traditional statistical analysis methods are highly uncertain, making it difficult to achieve early warning of agricultural drought.
Using machine learning methods, by collecting and preprocessing hydrological and meteorological data, using the extreme gradient boosting algorithm to identify key driving factors, a drought propagation time prediction model was constructed, including daily sliding accumulation calculation, inverse normal distribution transformation, and run theory to identify drought events and match drought event pairs. The key variables affecting the drought propagation time were screened out, and the XGBoost regression model was trained for prediction.
It has achieved accurate prediction of the transmission time of meteorological drought to agricultural drought at high resolution, can provide early warning of agricultural drought after the occurrence of meteorological drought events, and provide technical support for agricultural drought prediction and optimal allocation of water resources.
Smart Images
Figure CN120782035A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of drought disaster prediction, and in particular to a drought propagation time prediction method based on machine learning. Background Art
[0002] Drought is one of the most widespread and costly natural disasters in the world. Different types of drought are closely linked, with meteorological drought being a precursor to agricultural drought. Long-term rainfall deficiencies (meteorological droughts) lead to a sustained decline in soil moisture, which in turn affects vegetation growth and crop production, ultimately triggering agricultural drought. The transition from meteorological drought to agricultural drought is known as drought propagation, and drought propagation time is an important indicator of the lag between agricultural and meteorological droughts. Accurately predicting the propagation time from meteorological drought to agricultural drought is crucial for agricultural drought early warning and optimized water resource management.
[0003] Run theory and related threshold analysis methods are widely used in drought event identification. They can determine the onset, duration, and end time of drought events and are one of the main methods for estimating drought propagation time. This method uses the difference between the onset times of agricultural drought and meteorological drought events as the drought propagation time. Compared with traditional correlation analysis methods, it can more effectively reflect the event-dependent and time-varying characteristics of drought propagation time. However, existing research has mainly focused on statistical analysis of drought propagation time and lacks a practical drought propagation time prediction method, making it difficult to achieve early warning of agricultural drought based on meteorological drought information. Drought propagation time is not only affected by meteorological drought characteristics but also closely related to multiple factors such as antecedent soil moisture. Furthermore, the dominant factors influencing drought propagation time may vary significantly across regions and climatic conditions. Traditional statistical analysis methods struggle to accurately characterize the complex nonlinear relationships between drought propagation time and its multiple drivers, making it difficult to accurately identify the driving factors of drought propagation time, resulting in significant uncertainty in prediction results. With the rapid development of computer technology, machine learning methods have provided an effective means of characterizing the complex nonlinear relationships between hydrological and meteorological series, overcoming the limitations of traditional statistical analysis methods in this field. Therefore, how to use machine learning methods to extract the key driving factors of drought propagation time, accurately characterize the nonlinear relationship between them and propagation time, and construct a drought propagation time prediction model with adaptive capabilities, is still a key issue that needs to be urgently solved in current drought propagation research and prediction applications. Summary of the Invention
[0004] The technical problem to be solved by the present invention is the lack of drought propagation time prediction methods and the large uncertainty of the prediction results of the prediction model based on traditional statistical analysis methods, which makes it difficult to meet the actual needs of early warning of agricultural drought. In order to solve the above problems, a drought propagation time prediction method based on machine learning is provided.
[0005] The object of the present invention is achieved in the following manner: A method for predicting drought propagation time based on machine learning, the method comprising: S1: Collect and pre-process hydrometeorological data for drought index calculation; the hydrometeorological data include daily precipitation, potential evapotranspiration, and soil moisture; S2: Calculate cumulative precipitation, potential evapotranspiration, and soil moisture at a given target time scale using a daily sliding method; fit the probability distribution function of the cumulative data and perform an inverse normal distribution transformation on the cumulative probability to obtain meteorological drought and agricultural drought index series; S3: Use the run theory to identify meteorological drought and agricultural drought events, and statistically analyze the soil moisture status at the time of meteorological drought events, as well as the occurrence time, development duration, and intensity of meteorological drought events as potential factors affecting the propagation time of drought; S4: Match the identified meteorological drought and agricultural drought events and calculate the propagation time of each matched drought event pair; S5: Use the extreme gradient boosting algorithm to identify the key driving factors that affect the drought propagation time from the potential factors in step 3. Take the identified key driving factors as input and the drought propagation time in step 4 as the training target to train the drought propagation time prediction model based on the extreme gradient boosting algorithm to achieve the prediction of drought propagation time.
[0006] The S1 includes: resampling the collected data to a daily time resolution.
[0007] In S2, precipitation, potential evapotranspiration and soil moisture data are calculated on a daily sliding accumulation basis on a 30-day time scale, and a standardized precipitation evapotranspiration index (SPEI) series and a standardized soil moisture index (SSMI) series are constructed respectively.
[0008] The construction of the standardized precipitation evapotranspiration index SPEI sequence specifically includes: The GPP algorithm is used to calculate the cumulative probability of cumulative water deficit, which is the difference between cumulative precipitation and potential evapotranspiration. Since the cumulative water deficit on different dates presents its own unique probability characteristics, it is necessary to calculate the cumulative probability of cumulative water deficit on each date within the study period separately. The cumulative probability is calculated using formula (1): (1) Where, P is the cumulative probability, i is the order of the cumulative water deficit sequence in ascending order, and n is the length of the sample sequence; The calculated cumulative probability of cumulative water deficit is converted into the corresponding value in the standard normal distribution through equal probability transformation, which is SPEI: (2) Where Φ is the standard normal distribution function.
[0009] The S3 method includes: using run theory and -0.50 as a threshold to identify meteorological drought and agricultural drought events; defining drought duration as the number of consecutive days with a Standardized Precipitation Evaporation Index (SPEI) or Standardized Soil Moisture Index (SSMI) value below -0.5, and selecting 30 days as the drought duration threshold to identify a drought event; to avoid misdividing consecutive drought events into multiple shorter events due to short-term precipitation replenishment during the drought period, thereby affecting the statistics of drought duration, intensity and other characteristics, it is necessary to restore several "correlated drought" events into a single drought event; using 5 days as the threshold for merging "correlated drought" events, that is, merging two adjacent drought events with a time interval of no more than 5 days into a single drought event; The drought process is usually divided into a development stage and a recovery stage. The development stage refers to the period from the onset of the drought to the peak of the drought intensity, while the recovery stage refers to the period from the peak intensity of the drought to the end of the drought event. The soil moisture state at the time of the meteorological drought event, the characteristics of the development stage of the meteorological drought event, and the date of the meteorological drought event are selected as potential factors affecting the duration of drought propagation. Specifically, the soil moisture state IC at the time of the meteorological drought is represented by the SSMI value on the day of the meteorological drought event. The development duration DDD of the meteorological drought event is defined as the number of days from the onset of the drought event to the maximum intensity of the drought. The intensity DDI of the meteorological drought event during the development period is expressed as the average daily SPEI during the drought development stage. In addition, the date of the meteorological drought event is represented by the Day of the Year (DOY), which is the number of days accumulated starting from January 1 of the year.
[0010] S4 includes: before calculating the propagation time from the meteorological drought event to the agricultural drought event, it is necessary to match the agricultural drought event identified in step 3 with the meteorological drought event to construct a drought event pair; if the agricultural drought event and the meteorological drought event meet any of the following conditions, the two are considered to constitute a drought event pair: first, the agricultural drought event and the meteorological drought event overlap in time, and the agricultural drought occurs at most 10 days earlier than the meteorological drought event; second, the two types of drought events do not overlap in time, but the start time of the agricultural drought event is within 10 days after the end of the meteorological drought event; Based on the matching drought events, this embodiment defines the drought propagation time as the time difference between the occurrence time of the agricultural drought event and the occurrence time of the meteorological drought event: (3) Where, PT The drought propagation time, Ad st,i and Md st,i Respectively i The occurrence time of agricultural drought events and meteorological drought events in the matched drought event pairs.
[0011] In said S5: soil moisture state at the time of occurrence of meteorological drought event, duration of meteorological drought event, intensity during development of meteorological drought event and date of occurrence of meteorological drought event are selected as potential factors affecting drought propagation time; First, the normalization method was used to normalize the four potential influencing factors mentioned above. Then, the extreme gradient boosting algorithm was used to identify the main driving factors that have a significant impact on the drought propagation time: (4) Where min(x) and max(x) are the minimum and maximum values of the potential influencing factor x, respectively.
[0012] In S5, the extreme gradient boosting regression model was used to analyze the importance of each potential influencing factor, thereby screening out the variables that have the most significant impact on the drought propagation time. The specific process is as follows: First, the XGBoost regression model is initialized, and the soil moisture state at the time of the meteorological drought event, the duration of the meteorological drought event, the intensity during the development of the meteorological drought event, and the date of the meteorological drought event are input into the model as predictor variables. The drought propagation time is used as the target variable, and the XGBoost regression model is trained to capture the relationship between the variables. After model training is completed, the relative importance of each potential influencing factor is calculated, and based on this, the key variables affecting the drought transmission time are selected; the relative importance score of each potential influencing factor is calculated by calling the feature_importance() function in the Python language xgboost software package; According to the calculation results of the above relative importance, the soil moisture state at the time of the meteorological drought event, the development duration of the meteorological drought event and the date of occurrence of the meteorological drought event were selected as the final drought propagation time prediction factors.
[0013] In S5, after completing the key variable screening, the selected key variables are used as predictors to reconstruct the XGBoost regression model to predict the drought propagation time. The specific process is as follows: Dataset division: All samples are randomly divided into 80% training set and 20% test set, where the training set is used for model training and the test set is used for model evaluation; Model hyperparameter tuning: Based on the initial model construction, we focused on optimizing hyperparameters such as learning rate and maximum depth. We used a grid search method to find the optimal parameter combination for the XGBoost model on the training dataset. We also used 5-fold cross-validation to evaluate the robustness of the model on different data subsets. Prediction model evaluation: After determining the optimal hyperparameters, the prediction model is finally evaluated on the test set. By calculating the root mean square error (RMSE) and correlation coefficient (CC) indicators, the prediction accuracy and fitting effect of the model are comprehensively measured to verify the effectiveness of the model: (5) (6) Where, and are the predicted and observed drought propagation times, and are the average values of predicted and observed drought propagation time, respectively, and n is the number of samples.
[0014] The present invention utilizes machine learning algorithms to identify key drivers affecting drought propagation time and constructs a prediction model. This helps fully exploit the complex, nonlinear relationship between these drivers and propagation time, thereby improving the accuracy of drought propagation time prediction. This method can predict the propagation time from meteorological drought to agricultural drought at high resolution and daily scales based on information such as meteorological drought characteristics and initial soil moisture status. This allows for precise early warning of the onset of agricultural drought after a meteorological drought event occurs, providing technical support for agricultural drought prediction and the optimal allocation of water resources for drought relief. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 This is a flow chart of a drought propagation time prediction method based on machine learning proposed in the present invention.
[0016] Figure 2 is the importance score of each variable estimated by the extreme gradient boosting algorithm of the present invention.
[0017] Figure 3 This is a comparison chart of the propagation time predicted by the present invention and the actual propagation time. DETAILED DESCRIPTION
[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0019] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the present application. Unless otherwise defined, all technical and scientific terms used herein have the same technical meaning as commonly understood by one of ordinary skill in the art to which this application belongs.
[0020] The drought propagation time prediction method described in the present application has a flowchart as shown in Figure 1 The drought propagation time prediction method described in the present application has a flowchart as shown in Step 1: Collect and pre-process daily precipitation, potential evapotranspiration and soil moisture data for drought index calculation.
[0021] The present embodiment takes the Yangtze River Basin as the research area, and divides the Yangtze River Basin into four sub-areas (Yangtze River Basin, South China, North China and Northwest China). Figure 2 The research period is from 1980 to 2024. Daily precipitation, potential evapotranspiration and soil moisture data in the study area are collected. As an example, the daily precipitation, potential evapotranspiration and soil moisture data used here come from ERA5-Land (European Centre for Medium-Range Weather Forecasts Fifth Generation Reanalysis of the Global Atmosphere Land Dataset), which has a time resolution of hourly and a spatial resolution of 0.10°. In this embodiment, it is resampled to a daily time scale.
[0022] Step 2: Calculate the cumulative precipitation, potential evapotranspiration and soil moisture at a given target time scale using the daily sliding method; fit the probability distribution function of the cumulative data, and perform inverse normal distribution transformation on the cumulative probability to obtain the meteorological drought and agricultural drought index sequence.
[0023] In this embodiment, the standardized precipitation evapotranspiration index (SPEI) and the standardized soil moisture index (SSMI) are selected to identify meteorological drought and agricultural drought, respectively. To improve the time resolution of the drought propagation time identification results, this example selects a 30-day time scale to calculate the daily sliding cumulative precipitation, potential evapotranspiration and soil moisture data, and constructs the SPEI and SSMI sequences respectively. The calculation process of SPEI and SSMI is basically the same, the difference is that SSMI only depends on soil moisture data, while SPEI is based on precipitation and potential evapotranspiration data. Therefore, this embodiment only takes SPEI as an example to illustrate the specific calculation process of the drought index.
[0024] In order to avoid the assumption of the distribution function and reduce the calculation of the fitted parameter distribution, this embodiment uses the GPP (Gringorten plotting position) algorithm to calculate the cumulative probability of the cumulative water deficit (i.e. the difference between the cumulative precipitation and the potential evapotranspiration). Since the cumulative water deficit of different dates presents its own unique probability characteristics, it is necessary to calculate the cumulative probability of the cumulative water deficit of each date in the study period. For a certain date dFor example, extract all years between 1980-2021 on the date d The corresponding cumulative water deficit amount, and use formula (1) to calculate its cumulative probability: (1) In the formula, P is the cumulative probability, i is the order when the cumulative water deficit amount sequence is in ascending order, and n is the length of the sample sequence (42 in this embodiment).
[0025] Through equal probability transformation, the cumulative probability of the calculated cumulative water deficit amount is converted into the corresponding value in the standard normal distribution, which is the SPEI: (2) In the formula, Φ is the standard normal distribution function.
[0026] Step 3: Identify meteorological drought and agricultural drought events using run theory, and count the soil water content state when meteorological drought events occur, as well as the occurrence time, development duration, and development period intensity of meteorological drought events, as potential factors affecting drought transmission time.
[0027] Based on the meteorological drought index and agricultural drought index calculated in step 2, this embodiment uses run theory and takes -0.50 as the threshold to identify meteorological drought and agricultural drought events. Specifically, the duration of drought is defined as the number of consecutive days when the SPEI or SSMI value is below -0.5, and this embodiment selects 30 days as the threshold for drought duration to identify a drought event. To avoid the influence of short-term precipitation replenishment during drought on the statistical characteristics of drought duration and intensity, it is necessary to restore several "associated drought" events to a drought event. This embodiment uses 5 days as the threshold for merging "associated drought" events, that is, if the time interval between two adjacent drought events is not more than 5 days, they are merged into one drought event.
[0028] The drought process is usually divided into a development phase and a recovery phase: the drought development phase refers to a period of time from the beginning of the drought to the time when the drought intensity reaches a peak (i.e., the minimum value of the drought index during the drought), and the drought recovery phase refers to a period of time from the peak intensity of the drought to the end of the drought event. In this embodiment, the soil water content state at the time of the occurrence of the meteorological drought event, the characteristics of the development phase of the meteorological drought event, and the occurrence date of the meteorological drought event are selected as potential factors affecting the drought propagation time. Specifically, the soil water content state (IC) at the time of the occurrence of the meteorological drought event is represented by the SSMI value on the day of the occurrence of the meteorological drought event; the development duration of the meteorological drought event (DDD) is defined as the number of days from the beginning of the drought event to the time when the drought intensity reaches the maximum; and the intensity during the development phase of the meteorological drought event (DDI) is represented by the average value of the daily SPEI during the drought development phase. In addition, the occurrence date of the meteorological drought event is represented by the day of the year (DOY), i.e., the number of days accumulated from January 1 of the current year. For example, a meteorological drought event occurring on January 1 corresponds to a DOY of 1, and a meteorological drought event occurring on February 1 corresponds to a DOY of 32.
[0029] Step 4: Matching the identified meteorological drought and agricultural drought events, and calculating the propagation time of each matched drought event pair.
[0030] Before calculating the propagation time of the meteorological drought event to the agricultural drought event, the agricultural drought events and the meteorological drought events identified in step 3 need to be matched to construct drought event pairs. In this embodiment, if the agricultural drought event and the meteorological drought event satisfy any of the following conditions, they are considered to constitute a drought event pair: (1) the agricultural drought event and the meteorological drought event overlap in time, and the occurrence time of the agricultural drought is at most 10 days earlier than the meteorological drought event; (2) the two types of drought events do not overlap in time, but the start time of the agricultural drought event is within 10 days after the end of the meteorological drought event.
[0031] Based on the matched drought events, this embodiment defines the drought propagation time as the time difference between the occurrence time of the agricultural drought event and the occurrence time of the meteorological drought event: (3) In the formula, PT is the drought propagation time, Ad st,i and Md st,i are the occurrence times of the agricultural drought event and the meteorological drought event, respectively, in the i-th matched drought event pair. i
[0032] Step 5: Use the extreme gradient boosting algorithm to identify the key driving factors that affect the drought propagation time from the potential factors in step 3. Using the identified key driving factors as input and the drought propagation time in step 4 as the training target, train the drought propagation time prediction model based on the extreme gradient boosting algorithm to achieve the prediction of drought propagation time.
[0033] Before building a drought propagation time prediction model, it is first necessary to identify the key driving factors that influence drought propagation time. This example selects soil moisture at the time of the meteorological drought event, the duration of the meteorological drought event, the intensity of the meteorological drought event during its development, and the date of the meteorological drought event as potential factors influencing drought propagation time. Before identifying the key driving factors, a normalization method is first used to normalize the four potential influencing factors to eliminate the influence of unit and scale differences between the factors. Subsequently, the extreme gradient boosting algorithm is used to identify the main driving factors with a significant impact on drought propagation time: (4) Where min(x) and max(x) are the minimum and maximum values of the potential influencing factor x, respectively.
[0034] On this basis, the extreme gradient boosting regression model was used to analyze the importance of each potential influencing factor, thereby screening out the variables that have the most significant impact on the drought transmission time. The specific process is as follows: First, the XGBoost regression model is initialized, and the soil moisture state at the time of the meteorological drought event, the development duration of the meteorological drought event, the intensity during the development period of the meteorological drought event, and the date of occurrence of the meteorological drought event are input into the model as predictor variables. The drought propagation time is used as the target variable, and the XGBoost regression model is trained to capture the relationship between the variables.
[0035] After the model training is completed, the relative importance of each potential influencing factor is calculated, and the key variables affecting the drought propagation time are selected accordingly. Specifically, this embodiment calculates the relative importance score of each potential influencing factor by calling the feature_importance() function in the Python language xgboost software package ( Figure 2 The average relative importance scores of soil moisture status at the time of meteorological drought event, intensity during the development period of meteorological drought event, duration of meteorological drought event and date of occurrence of meteorological drought event were 0.42, 0.25, 0.12 and 0.22, respectively.
[0036] Based on the calculation results of the above relative importance, this embodiment selects the soil moisture state at the time of the meteorological drought event, the development duration of the meteorological drought event, and the date of occurrence of the meteorological drought event as the final drought propagation time prediction factors, thereby eliminating variables that contribute less to the prediction and reducing the complexity of the model.
[0037] After completing the key variable screening, this example uses the screened key variables as predictors to rebuild the XGBoost regression model to predict the drought spread time. The specific process is as follows: Dataset partitioning: All samples are randomly divided into 80% training set and 20% test set, where the training set is used for model training and the test set is used for model evaluation.
[0038] Model hyperparameter tuning. Based on the initial model construction, we focused on optimizing hyperparameters such as the learning rate and maximum depth. We used a grid search method to find the optimal parameter combination for the XGBoost model on the training dataset. We also used 5-fold cross-validation to evaluate the robustness of the model on different data subsets, ensuring consistent performance across different data subsets and improving the model's generalization performance.
[0039] Prediction model evaluation. After determining the optimal hyperparameters, the prediction model is finally evaluated on the test set. By calculating indicators such as the root mean square error (RMSE) and correlation coefficient (CC), the model's prediction accuracy and fit are comprehensively measured to verify the effectiveness of the model: (5) (6) Where, and are the predicted and observed drought propagation times, and are the average values of predicted and observed drought propagation time, respectively, and n is the number of samples.
[0040] To intuitively evaluate the prediction ability of the model, this example draws a scatter plot of the predicted propagation time and the observed propagation time during the test period ( Figure 3 ).
[0041] As can be seen, the data points for predicted and observed propagation times are mostly concentrated near the 1:1 diagonal, indicating a high degree of agreement between the predicted and observed values. The correlation coefficients for the predicted and observed propagation times for the four subregions ranged from 0.72 to 0.82, and the root mean square error (RMS) ranged from 7.1 to 17.1 days, demonstrating the model's high prediction accuracy. Therefore, this method provides an effective method for predicting drought propagation time, enabling high-resolution daily-scale prediction of the propagation time from meteorological drought to agricultural drought. This method can effectively address the inaccurate prediction of agricultural drought onset times caused by a lack of drought propagation time prediction methods, providing important support for improving regional scientific drought prevention and effective drought relief capabilities, and possesses significant scientific value and application prospects.
[0042] The above is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several changes and improvements can be made without departing from the overall concept of the present invention, and these should also be regarded as the scope of protection of the present invention.
Claims
1. A method for predicting drought spread time based on machine learning, characterized by: The method comprises: S1: Collect and pre-process hydrometeorological data for drought index calculation; the hydrometeorological data include daily precipitation, potential evapotranspiration, and soil moisture; S2: Calculate cumulative precipitation, potential evapotranspiration, and soil moisture at a given target time scale using a daily sliding method; fit the probability distribution function of the cumulative data and perform an inverse normal distribution transformation on the cumulative probability to obtain meteorological drought and agricultural drought index series; S3: Use the run theory to identify meteorological drought and agricultural drought events, and statistically analyze the soil moisture status at the time of meteorological drought events, as well as the occurrence time, development duration, and intensity of meteorological drought events as potential factors affecting the propagation time of drought; S4: Match the identified meteorological drought and agricultural drought events and calculate the propagation time of each matched drought event pair; S5: Use the extreme gradient boosting algorithm to identify the key driving factors that affect the drought propagation time from the potential factors in step 3. Take the identified key driving factors as input and the drought propagation time in step 4 as the training target to train the drought propagation time prediction model based on the extreme gradient boosting algorithm to achieve the prediction of drought propagation time.
2. The drought propagation time prediction method based on machine learning according to claim 1, characterized in that: The S1 includes: resampling the collected data to a daily time resolution.
3. The drought propagation time prediction method based on machine learning according to claim 1, characterized in that: In S2, precipitation, potential evapotranspiration and soil moisture data are calculated on a daily sliding accumulation basis on a 30-day time scale, and a standardized precipitation evapotranspiration index (SPEI) series and a standardized soil moisture index (SSMI) series are constructed respectively.
4. The method for predicting drought propagation time based on machine learning according to claim 3, characterized in that: The construction of the standardized precipitation evapotranspiration index SPEI sequence specifically includes: The GPP algorithm is used to calculate the cumulative probability of cumulative water deficit, which is the difference between cumulative precipitation and potential evapotranspiration. Since the cumulative water deficit on different dates presents its own unique probability characteristics, it is necessary to calculate the cumulative probability of cumulative water deficit on each date within the study period separately. The cumulative probability is calculated using formula (1): （1） Where, P is the cumulative probability, i is the order of the cumulative water deficit sequence in ascending order, and n is the length of the sample sequence; The calculated cumulative probability of cumulative water deficit is converted into the corresponding value in the standard normal distribution through equal probability transformation, which is SPEI: （2） Where Φ is the standard normal distribution function.
5. The method for predicting drought propagation time based on machine learning according to claim 1, characterized in that: The S3 method includes: using run theory and a threshold of -0.50 to identify meteorological and agricultural drought events; defining drought duration as the number of consecutive days with a Standardized Precipitation Evapotranspiration Index (SPEI) or Standardized Soil Moisture Index (SSMI) value below -0.5, and selecting 30 days as the drought duration threshold to identify a drought event; to avoid misclassifying consecutive drought events into multiple shorter events due to short-term precipitation replenishment during the drought, thereby affecting the statistical characteristics of drought duration and intensity, it is necessary to restore multiple "related drought" events into a single drought event; using 5 days as the threshold for merging "related drought" events, that is, merging two adjacent drought events with a time interval of no more than 5 days into a single drought event; The drought process is usually divided into a development stage and a recovery stage. The development stage refers to the period from the onset of the drought to the peak of the drought intensity, while the recovery stage refers to the period from the peak intensity of the drought to the end of the drought event. The soil moisture state at the time of the meteorological drought event, the characteristics of the development stage of the meteorological drought event, and the date of the meteorological drought event are selected as potential factors affecting the duration of drought propagation. Specifically, the soil moisture state IC at the time of the meteorological drought is represented by the SSMI value on the day of the meteorological drought event. The development duration DDD of the meteorological drought event is defined as the number of days from the onset of the drought event to the maximum intensity of the drought. The intensity DDI of the meteorological drought event during the development period is expressed as the average daily SPEI during the drought development stage. In addition, the date of the meteorological drought event is represented by the Day of the Year (DOY), which is the number of days accumulated starting from January 1 of the year.
6. The method for predicting drought propagation time based on machine learning according to claim 1, characterized in that: S4 includes: before calculating the propagation time from the meteorological drought event to the agricultural drought event, it is necessary to match the agricultural drought event identified in step 3 with the meteorological drought event to construct a drought event pair; if the agricultural drought event and the meteorological drought event meet any of the following conditions, the two are considered to constitute a drought event pair: first, the agricultural drought event and the meteorological drought event overlap in time, and the agricultural drought occurs at most 10 days earlier than the meteorological drought event; second, the two types of drought events do not overlap in time, but the start time of the agricultural drought event is within 10 days after the end of the meteorological drought event; Based on the matching drought events, this embodiment defines the drought propagation time as the time difference between the occurrence time of the agricultural drought event and the occurrence time of the meteorological drought event: （3） Where, PT The drought propagation time, Ad st,i and Md st,i Respectively i The occurrence time of agricultural drought events and meteorological drought events in the matched drought event pairs.
7. The method for predicting drought propagation time based on machine learning according to claim 1, characterized in that: In said S5: soil moisture state at the time of occurrence of meteorological drought event, duration of meteorological drought event, intensity during development of meteorological drought event and date of occurrence of meteorological drought event are selected as potential factors affecting drought propagation time; First, the normalization method was used to normalize the four potential influencing factors mentioned above. Then, the extreme gradient boosting algorithm was used to identify the main driving factors that have a significant impact on the drought propagation time: （4） Where min(x) and max(x) are the minimum and maximum values of the potential influencing factor x, respectively.
8. The method for predicting drought propagation time based on machine learning according to claim 1, characterized in that: The S5 The extreme gradient boosting regression model was used to analyze the importance of each potential influencing factor, thereby screening out the variables with the most significant impact on drought propagation time. The specific process is as follows: First, the XGBoost regression model is initialized, and the soil moisture state at the time of the meteorological drought event, the duration of the meteorological drought event, the intensity during the development of the meteorological drought event, and the date of the meteorological drought event are input into the model as predictor variables. The drought propagation time is used as the target variable, and the XGBoost regression model is trained to capture the relationship between the variables. After model training is completed, the relative importance of each potential influencing factor is calculated, and based on this, the key variables affecting the drought transmission time are selected; the relative importance score of each potential influencing factor is calculated by calling the feature_importance() function in the Python language xgboost software package; According to the calculation results of the above relative importance, the soil moisture state at the time of the meteorological drought event, the development duration of the meteorological drought event and the date of occurrence of the meteorological drought event were selected as the final drought propagation time prediction factors.
9. The method for predicting drought propagation time based on machine learning according to claim 1, characterized in that: In said S5: After completing the key variable screening, the selected key variables were used as predictors to rebuild the XGBoost regression model to predict the drought spread time. The specific process is as follows: Dataset division: All samples are randomly divided into 80% training set and 20% test set, where the training set is used for model training and the test set is used for model evaluation; Model hyperparameter tuning: Based on the initial model construction, we focused on optimizing hyperparameters such as learning rate and maximum depth. We used a grid search method to find the optimal parameter combination for the XGBoost model on the training dataset. We also used 5-fold cross-validation to evaluate the robustness of the model on different data subsets. Prediction model evaluation: After determining the optimal hyperparameters, the prediction model is finally evaluated on the test set. By calculating the root mean square error (RMSE) and correlation coefficient (CC) indicators, the prediction accuracy and fitting effect of the model are comprehensively measured to verify the effectiveness of the model: （5） （6） Where, and are the predicted and observed drought propagation times, and are the average values of predicted and observed drought propagation time, respectively, and n is the number of samples.